DeeBLiF

Zhengyu Zhang, Shishun Tian, Wenbin Zou, Luce Morin, and Lu Zhang.

Official PyTorch code for our ICIP2022 paper "DeeBLiF: Deep blind light field image quality assessment by extracting angular and spatial information". Please refer to our paper for details.

Note: We first convert the dataset into h5 files in MATLAB and then train/test the model in PYTHON.

Hope our work is helpful to you :)

Requirements

• PyTorch 1.7.1
• python 3.8

Installation

Download this repository:

    $ git clone https://github.com/ZhengyuZhang96/DeeBLiF.git

Generate Dataset in MATLAB

Take the Win5-LID dataset for instance, download the .m file on Google drive or Baidu drive (code: INSA), convert the dataset into h5 files, and then put them into './DeeBLiF/Datasets/Win5_160x160/...':

    $ ./DeeBLiF/Datasets/Generateh5_for_Win5_Dataset.m

or you can directly download the generated h5 files on Google drive or Baidu drive (code: INSA).

Train

Train the model from scratch:

    $ python Train.py  --trainset_dir ./Datasets/Win5_160x160/

Test overall performance

Reproduce the performance in the paper: download our pre-trained models on Google drive or Baidu drive (code: INSA) and put them into './DeeBLiF/PreTrainedModels/Win5/...'.

    $ python Test.py

Test individual distortion type performance

Test the performance of individual distortion type by the following script.

    $ python Test_Dist.py

Results

Our paper only provides the experimental results of the overall performance on the Win5-LID dataset, here we additionally provide the individual distortion type performance of the Win5-LID dataset, and the individual distortion type performance and overall performance of the NBU-LF1.0 and SHU datasets. Alternatively, you can reproduce these performances using the h5 results we provide in './DeeBLiF/Results/...'.

Win5-LID dataset:

Distortion types	PLCC	SROCC	KROCC	RMSE
HEVC	0.9389	0.9103	0.7988	0.3406
JPEG2000	0.9254	0.8686	0.7508	0.3257
LN	0.9021	0.7914	0.6548	0.2964
NN	0.9207	0.8628	0.7382	0.2701
Overall	0.8427	0.8186	0.6502	0.5160

NBU-LF1.0 dataset:

Distortion types	PLCC	SROCC	KROCC	RMSE
NN	0.9610	0.9168	0.8100	0.1843
BI	0.9499	0.8986	0.7918	0.2736
EPICNN	0.9395	0.8027	0.6899	0.2283
Zhang	0.6659	0.5832	0.5003	0.4365
VDSR	0.9487	0.9062	0.8042	0.2614
Overall	0.8583	0.8229	0.6515	0.4588

SHU dataset:

Distortion types	PLCC	SROCC	KROCC	RMSE
GAUSS	0.9556	0.9507	0.8609	0.2238
JPEG2000	0.9031	0.8980	0.7962	0.1620
JPEG	0.9804	0.9567	0.8641	0.2040
Motion Blur	0.9676	0.9474	0.8516	0.2099
White Noise	0.9553	0.9527	0.8709	0.2832
Overall	0.9548	0.9419	0.8149	0.3185

Citation

If you find this work helpful, please consider citing:

@inproceedings{zhang2022deeblif,
  title        = {Deeblif: Deep Blind Light Field Image Quality Assessment by Extracting Angular and Spatial Information},
  author       = {Zhang, Zhengyu and Tian, Shishun and Zou, Wenbin and Morin, Luce and Zhang, Lu},
  booktitle    = {2022 IEEE International Conference on Image Processing (ICIP)},
  pages        = {2266--2270},
  year         = {2022},
  organization = {IEEE}
}

Correction

In our paper, we claim that we use "K-fold cross-validation" strategy to conducet the experiments. However, it should actually be "Leave-two-fold-out cross-validation". We sincerely apologize for any confusion or inconvenience caused by this wrong expression.

Contact

Welcome to raise issues or email to [email protected] for any question regarding this work.